2 * Copyright (C) 1995-2010 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief This file implements the x87 support and virtual to stack
23 * register translation for the ia32 backend.
24 * @author Michael Beck
33 #include "iredges_t.h"
48 #include "bearch_ia32_t.h"
49 #include "ia32_new_nodes.h"
50 #include "gen_ia32_new_nodes.h"
51 #include "gen_ia32_regalloc_if.h"
53 #include "ia32_architecture.h"
55 /** the debug handle */
56 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
58 /* Forward declaration. */
59 typedef struct x87_simulator x87_simulator;
62 * An exchange template.
63 * Note that our virtual functions have the same inputs
64 * and attributes as the real ones, so we can simple exchange
66 * Further, x87 supports inverse instructions, so we can handle them.
68 typedef struct exchange_tmpl {
69 ir_op *normal_op; /**< the normal one */
70 ir_op *reverse_op; /**< the reverse one if exists */
71 ir_op *normal_pop_op; /**< the normal one with tos pop */
72 ir_op *reverse_pop_op; /**< the reverse one with tos pop */
76 * An entry on the simulated x87 stack.
78 typedef struct st_entry {
79 int reg_idx; /**< the virtual register index of this stack value */
80 ir_node *node; /**< the node that produced this value */
86 typedef struct x87_state {
87 st_entry st[N_ia32_st_REGS]; /**< the register stack */
88 int depth; /**< the current stack depth */
89 x87_simulator *sim; /**< The simulator. */
92 /** An empty state, used for blocks without fp instructions. */
93 static x87_state empty = { { {0, NULL}, }, 0, NULL };
96 * Return values of the instruction simulator functions.
99 NO_NODE_ADDED = 0, /**< No node that needs simulation was added. */
100 NODE_ADDED = 1 /**< A node that must be simulated was added by the simulator
101 in the schedule AFTER the current node. */
105 * The type of an instruction simulator function.
107 * @param state the x87 state
108 * @param n the node to be simulated
110 * @return NODE_ADDED if a node was added AFTER n in schedule that MUST be
112 * NO_NODE_ADDED otherwise
114 typedef int (*sim_func)(x87_state *state, ir_node *n);
117 * A block state: Every block has a x87 state at the beginning and at the end.
119 typedef struct blk_state {
120 x87_state *begin; /**< state at the begin or NULL if not assigned */
121 x87_state *end; /**< state at the end or NULL if not assigned */
124 /** liveness bitset for vfp registers. */
125 typedef unsigned char vfp_liveness;
130 struct x87_simulator {
131 struct obstack obst; /**< An obstack for fast allocating. */
132 pmap *blk_states; /**< Map blocks to states. */
133 be_lv_t *lv; /**< intrablock liveness. */
134 vfp_liveness *live; /**< Liveness information. */
135 unsigned n_idx; /**< The cached get_irg_last_idx() result. */
136 waitq *worklist; /**< Worklist of blocks that must be processed. */
140 * Returns the current stack depth.
142 * @param state the x87 state
144 * @return the x87 stack depth
146 static int x87_get_depth(const x87_state *state)
151 static st_entry *x87_get_entry(x87_state *const state, int const pos)
153 assert(0 <= pos && pos < state->depth);
154 return &state->st[N_ia32_st_REGS - state->depth + pos];
158 * Return the virtual register index at st(pos).
160 * @param state the x87 state
161 * @param pos a stack position
163 * @return the vfp register index that produced the value at st(pos)
165 static int x87_get_st_reg(const x87_state *state, int pos)
167 return x87_get_entry((x87_state*)state, pos)->reg_idx;
172 * Dump the stack for debugging.
174 * @param state the x87 state
176 static void x87_dump_stack(const x87_state *state)
178 for (int i = state->depth; i-- != 0;) {
179 st_entry const *const entry = x87_get_entry((x87_state*)state, i);
180 DB((dbg, LEVEL_2, "vf%d(%+F) ", entry->reg_idx, entry->node));
182 DB((dbg, LEVEL_2, "<-- TOS\n"));
184 #endif /* DEBUG_libfirm */
187 * Set a virtual register to st(pos).
189 * @param state the x87 state
190 * @param reg_idx the vfp register index that should be set
191 * @param node the IR node that produces the value of the vfp register
192 * @param pos the stack position where the new value should be entered
194 static void x87_set_st(x87_state *state, int reg_idx, ir_node *node, int pos)
196 st_entry *const entry = x87_get_entry(state, pos);
197 entry->reg_idx = reg_idx;
200 DB((dbg, LEVEL_2, "After SET_REG: "));
201 DEBUG_ONLY(x87_dump_stack(state);)
205 * Set the tos virtual register.
207 * @param state the x87 state
208 * @param reg_idx the vfp register index that should be set
209 * @param node the IR node that produces the value of the vfp register
211 static void x87_set_tos(x87_state *state, int reg_idx, ir_node *node)
213 x87_set_st(state, reg_idx, node, 0);
217 * Swap st(0) with st(pos).
219 * @param state the x87 state
220 * @param pos the stack position to change the tos with
222 static void x87_fxch(x87_state *state, int pos)
224 st_entry *const a = x87_get_entry(state, pos);
225 st_entry *const b = x87_get_entry(state, 0);
226 st_entry const t = *a;
230 DB((dbg, LEVEL_2, "After FXCH: "));
231 DEBUG_ONLY(x87_dump_stack(state);)
235 * Convert a virtual register to the stack index.
237 * @param state the x87 state
238 * @param reg_idx the register vfp index
240 * @return the stack position where the register is stacked
241 * or -1 if the virtual register was not found
243 static int x87_on_stack(const x87_state *state, int reg_idx)
245 for (int i = 0; i < state->depth; ++i) {
246 if (x87_get_st_reg(state, i) == reg_idx)
253 * Push a virtual Register onto the stack, double pushed allowed.
255 * @param state the x87 state
256 * @param reg_idx the register vfp index
257 * @param node the node that produces the value of the vfp register
259 static void x87_push_dbl(x87_state *state, int reg_idx, ir_node *node)
261 assert(state->depth < N_ia32_st_REGS && "stack overrun");
264 st_entry *const entry = x87_get_entry(state, 0);
265 entry->reg_idx = reg_idx;
268 DB((dbg, LEVEL_2, "After PUSH: ")); DEBUG_ONLY(x87_dump_stack(state);)
272 * Push a virtual Register onto the stack, double pushes are NOT allowed.
274 * @param state the x87 state
275 * @param reg_idx the register vfp index
276 * @param node the node that produces the value of the vfp register
278 static void x87_push(x87_state *state, int reg_idx, ir_node *node)
280 assert(x87_on_stack(state, reg_idx) == -1 && "double push");
282 x87_push_dbl(state, reg_idx, node);
286 * Pop a virtual Register from the stack.
288 * @param state the x87 state
290 static void x87_pop(x87_state *state)
292 assert(state->depth > 0 && "stack underrun");
296 DB((dbg, LEVEL_2, "After POP: ")); DEBUG_ONLY(x87_dump_stack(state);)
300 * Empty the fpu stack
302 * @param state the x87 state
304 static void x87_emms(x87_state *state)
310 * Returns the block state of a block.
312 * @param sim the x87 simulator handle
313 * @param block the current block
315 * @return the block state
317 static blk_state *x87_get_bl_state(x87_simulator *sim, ir_node *block)
319 blk_state *res = pmap_get(blk_state, sim->blk_states, block);
322 res = OALLOC(&sim->obst, blk_state);
326 pmap_insert(sim->blk_states, block, res);
335 * @param sim the x87 simulator handle
336 * @param src the x87 state that will be cloned
338 * @return a cloned copy of the src state
340 static x87_state *x87_clone_state(x87_simulator *sim, const x87_state *src)
342 x87_state *const res = OALLOC(&sim->obst, x87_state);
348 * Patch a virtual instruction into a x87 one and return
349 * the node representing the result value.
351 * @param n the IR node to patch
352 * @param op the x87 opcode to patch in
354 static ir_node *x87_patch_insn(ir_node *n, ir_op *op)
356 ir_mode *mode = get_irn_mode(n);
361 if (mode == mode_T) {
362 /* patch all Proj's */
363 foreach_out_edge(n, edge) {
364 ir_node *proj = get_edge_src_irn(edge);
366 mode = get_irn_mode(proj);
367 if (mode_is_float(mode)) {
369 set_irn_mode(proj, ia32_reg_classes[CLASS_ia32_st].mode);
373 } else if (mode_is_float(mode))
374 set_irn_mode(n, ia32_reg_classes[CLASS_ia32_st].mode);
379 * Returns the first Proj of a mode_T node having a given mode.
381 * @param n the mode_T node
382 * @param m the desired mode of the Proj
383 * @return The first Proj of mode @p m found or NULL.
385 static ir_node *get_irn_Proj_for_mode(ir_node *n, ir_mode *m)
387 assert(get_irn_mode(n) == mode_T && "Need mode_T node");
389 foreach_out_edge(n, edge) {
390 ir_node *proj = get_edge_src_irn(edge);
391 if (get_irn_mode(proj) == m)
399 * Wrap the arch_* function here so we can check for errors.
401 static inline const arch_register_t *x87_get_irn_register(const ir_node *irn)
403 const arch_register_t *res = arch_get_irn_register(irn);
405 assert(res->reg_class == &ia32_reg_classes[CLASS_ia32_vfp]);
409 static inline const arch_register_t *x87_irn_get_register(const ir_node *irn,
412 const arch_register_t *res = arch_get_irn_register_out(irn, pos);
414 assert(res->reg_class == &ia32_reg_classes[CLASS_ia32_vfp]);
418 static inline const arch_register_t *get_st_reg(int index)
420 return &ia32_registers[REG_ST0 + index];
424 * Create a fxch node before another node.
426 * @param state the x87 state
427 * @param n the node after the fxch
428 * @param pos exchange st(pos) with st(0)
432 static ir_node *x87_create_fxch(x87_state *state, ir_node *n, int pos)
434 x87_fxch(state, pos);
436 ir_node *const block = get_nodes_block(n);
437 ir_node *const fxch = new_bd_ia32_fxch(NULL, block);
438 ia32_x87_attr_t *const attr = get_ia32_x87_attr(fxch);
439 attr->x87[0] = get_st_reg(pos);
440 attr->x87[2] = get_st_reg(0);
444 sched_add_before(n, fxch);
445 DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(fxch), attr->x87[0]->name, attr->x87[2]->name));
449 /* -------------- x87 perm --------------- */
452 * Calculate the necessary permutations to reach dst_state.
454 * These permutations are done with fxch instructions and placed
455 * at the end of the block.
457 * Note that critical edges are removed here, so we need only
458 * a shuffle if the current block has only one successor.
460 * @param block the current block
461 * @param state the current x87 stack state, might be modified
462 * @param dst_state destination state
466 static x87_state *x87_shuffle(ir_node *block, x87_state *state, const x87_state *dst_state)
468 int i, n_cycles, k, ri;
469 unsigned cycles[4], all_mask;
470 char cycle_idx[4][8];
472 assert(state->depth == dst_state->depth);
474 /* Some mathematics here:
475 * If we have a cycle of length n that includes the tos,
476 * we need n-1 exchange operations.
477 * We can always add the tos and restore it, so we need
478 * n+1 exchange operations for a cycle not containing the tos.
479 * So, the maximum of needed operations is for a cycle of 7
480 * not including the tos == 8.
481 * This is the same number of ops we would need for using stores,
482 * so exchange is cheaper (we save the loads).
483 * On the other hand, we might need an additional exchange
484 * in the next block to bring one operand on top, so the
485 * number of ops in the first case is identical.
486 * Further, no more than 4 cycles can exists (4 x 2). */
487 all_mask = (1 << (state->depth)) - 1;
489 for (n_cycles = 0; all_mask; ++n_cycles) {
490 int src_idx, dst_idx;
492 /* find the first free slot */
493 for (i = 0; i < state->depth; ++i) {
494 if (all_mask & (1 << i)) {
495 all_mask &= ~(1 << i);
497 /* check if there are differences here */
498 if (x87_get_st_reg(state, i) != x87_get_st_reg(dst_state, i))
504 /* no more cycles found */
509 cycles[n_cycles] = (1 << i);
510 cycle_idx[n_cycles][k++] = i;
511 for (src_idx = i; ; src_idx = dst_idx) {
512 dst_idx = x87_on_stack(dst_state, x87_get_st_reg(state, src_idx));
514 if ((all_mask & (1 << dst_idx)) == 0)
517 cycle_idx[n_cycles][k++] = dst_idx;
518 cycles[n_cycles] |= (1 << dst_idx);
519 all_mask &= ~(1 << dst_idx);
521 cycle_idx[n_cycles][k] = -1;
525 /* no permutation needed */
529 /* Hmm: permutation needed */
530 DB((dbg, LEVEL_2, "\n%+F needs permutation: from\n", block));
531 DEBUG_ONLY(x87_dump_stack(state);)
532 DB((dbg, LEVEL_2, " to\n"));
533 DEBUG_ONLY(x87_dump_stack(dst_state);)
537 DB((dbg, LEVEL_2, "Need %d cycles\n", n_cycles));
538 for (ri = 0; ri < n_cycles; ++ri) {
539 DB((dbg, LEVEL_2, " Ring %d:\n ", ri));
540 for (k = 0; cycle_idx[ri][k] != -1; ++k)
541 DB((dbg, LEVEL_2, " st%d ->", cycle_idx[ri][k]));
542 DB((dbg, LEVEL_2, "\n"));
547 * Find the place node must be insert.
548 * We have only one successor block, so the last instruction should
551 ir_node *const before = sched_last(block);
552 assert(is_cfop(before));
554 /* now do the permutations */
555 for (ri = 0; ri < n_cycles; ++ri) {
556 if ((cycles[ri] & 1) == 0) {
557 /* this cycle does not include the tos */
558 x87_create_fxch(state, before, cycle_idx[ri][0]);
560 for (k = 1; cycle_idx[ri][k] != -1; ++k) {
561 x87_create_fxch(state, before, cycle_idx[ri][k]);
563 if ((cycles[ri] & 1) == 0) {
564 /* this cycle does not include the tos */
565 x87_create_fxch(state, before, cycle_idx[ri][0]);
572 * Create a fpush before node n.
574 * @param state the x87 state
575 * @param n the node after the fpush
576 * @param pos push st(pos) on stack
577 * @param val the value to push
579 static void x87_create_fpush(x87_state *state, ir_node *n, int pos, ir_node *const val)
581 arch_register_t const *const out = x87_get_irn_register(val);
582 x87_push_dbl(state, arch_register_get_index(out), val);
584 ir_node *const fpush = new_bd_ia32_fpush(NULL, get_nodes_block(n));
585 ia32_x87_attr_t *const attr = get_ia32_x87_attr(fpush);
586 attr->x87[0] = get_st_reg(pos);
587 attr->x87[2] = get_st_reg(0);
590 sched_add_before(n, fpush);
592 DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(fpush), attr->x87[0]->name, attr->x87[2]->name));
596 * Create a fpop before node n.
598 * @param state the x87 state
599 * @param n the node after the fpop
600 * @param num pop 1 or 2 values
602 * @return the fpop node
604 static ir_node *x87_create_fpop(x87_state *state, ir_node *n, int num)
606 ir_node *fpop = NULL;
607 ia32_x87_attr_t *attr;
612 if (ia32_cg_config.use_ffreep)
613 fpop = new_bd_ia32_ffreep(NULL, get_nodes_block(n));
615 fpop = new_bd_ia32_fpop(NULL, get_nodes_block(n));
616 attr = get_ia32_x87_attr(fpop);
617 attr->x87[0] = get_st_reg(0);
618 attr->x87[1] = get_st_reg(0);
619 attr->x87[2] = get_st_reg(0);
622 sched_add_before(n, fpop);
623 DB((dbg, LEVEL_1, "<<< %s %s\n", get_irn_opname(fpop), attr->x87[0]->name));
628 /* --------------------------------- liveness ------------------------------------------ */
631 * The liveness transfer function.
632 * Updates a live set over a single step from a given node to its predecessor.
633 * Everything defined at the node is removed from the set, the uses of the node get inserted.
635 * @param irn The node at which liveness should be computed.
636 * @param live The bitset of registers live before @p irn. This set gets modified by updating it to
637 * the registers live after irn.
639 * @return The live bitset.
641 static vfp_liveness vfp_liveness_transfer(ir_node *irn, vfp_liveness live)
644 const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
646 if (get_irn_mode(irn) == mode_T) {
647 foreach_out_edge(irn, edge) {
648 ir_node *proj = get_edge_src_irn(edge);
650 if (arch_irn_consider_in_reg_alloc(cls, proj)) {
651 const arch_register_t *reg = x87_get_irn_register(proj);
652 live &= ~(1 << arch_register_get_index(reg));
655 } else if (arch_irn_consider_in_reg_alloc(cls, irn)) {
656 const arch_register_t *reg = x87_get_irn_register(irn);
657 live &= ~(1 << arch_register_get_index(reg));
660 for (i = 0, n = get_irn_arity(irn); i < n; ++i) {
661 ir_node *op = get_irn_n(irn, i);
663 if (mode_is_float(get_irn_mode(op)) &&
664 arch_irn_consider_in_reg_alloc(cls, op)) {
665 const arch_register_t *reg = x87_get_irn_register(op);
666 live |= 1 << arch_register_get_index(reg);
673 * Put all live virtual registers at the end of a block into a bitset.
675 * @param sim the simulator handle
676 * @param bl the block
678 * @return The live bitset at the end of this block
680 static vfp_liveness vfp_liveness_end_of_block(x87_simulator *sim, const ir_node *block)
682 vfp_liveness live = 0;
683 const arch_register_class_t *cls = &ia32_reg_classes[CLASS_ia32_vfp];
684 const be_lv_t *lv = sim->lv;
686 be_lv_foreach(lv, block, be_lv_state_end, node) {
687 const arch_register_t *reg;
688 if (!arch_irn_consider_in_reg_alloc(cls, node))
691 reg = x87_get_irn_register(node);
692 live |= 1 << arch_register_get_index(reg);
698 /** get the register mask from an arch_register */
699 #define REGMASK(reg) (1 << (arch_register_get_index(reg)))
702 * Return a bitset of argument registers which are live at the end of a node.
704 * @param sim the simulator handle
705 * @param pos the node
706 * @param kill kill mask for the output registers
708 * @return The live bitset.
710 static unsigned vfp_live_args_after(x87_simulator *sim, const ir_node *pos, unsigned kill)
712 unsigned idx = get_irn_idx(pos);
714 assert(idx < sim->n_idx);
715 return sim->live[idx] & ~kill;
719 * Calculate the liveness for a whole block and cache it.
721 * @param sim the simulator handle
722 * @param block the block
724 static void update_liveness(x87_simulator *sim, ir_node *block)
726 vfp_liveness live = vfp_liveness_end_of_block(sim, block);
729 /* now iterate through the block backward and cache the results */
730 sched_foreach_reverse(block, irn) {
731 /* stop at the first Phi: this produces the live-in */
735 idx = get_irn_idx(irn);
736 sim->live[idx] = live;
738 live = vfp_liveness_transfer(irn, live);
740 idx = get_irn_idx(block);
741 sim->live[idx] = live;
745 * Returns true if a register is live in a set.
747 * @param reg_idx the vfp register index
748 * @param live a live bitset
750 #define is_vfp_live(reg_idx, live) ((live) & (1 << (reg_idx)))
754 * Dump liveness info.
756 * @param live the live bitset
758 static void vfp_dump_live(vfp_liveness live)
762 DB((dbg, LEVEL_2, "Live after: "));
763 for (i = 0; i < 8; ++i) {
764 if (live & (1 << i)) {
765 DB((dbg, LEVEL_2, "vf%d ", i));
768 DB((dbg, LEVEL_2, "\n"));
770 #endif /* DEBUG_libfirm */
772 /* --------------------------------- simulators ---------------------------------------- */
775 * Simulate a virtual binop.
777 * @param state the x87 state
778 * @param n the node that should be simulated (and patched)
779 * @param tmpl the template containing the 4 possible x87 opcodes
781 * @return NO_NODE_ADDED
783 static int sim_binop(x87_state *state, ir_node *n, const exchange_tmpl *tmpl)
785 int op2_idx = 0, op1_idx;
786 int out_idx, do_pop = 0;
787 ia32_x87_attr_t *attr;
789 ir_node *patched_insn;
791 x87_simulator *sim = state->sim;
792 ir_node *op1 = get_irn_n(n, n_ia32_binary_left);
793 ir_node *op2 = get_irn_n(n, n_ia32_binary_right);
794 const arch_register_t *op1_reg = x87_get_irn_register(op1);
795 const arch_register_t *op2_reg = x87_get_irn_register(op2);
796 const arch_register_t *out = x87_irn_get_register(n, pn_ia32_res);
797 int reg_index_1 = arch_register_get_index(op1_reg);
798 int reg_index_2 = arch_register_get_index(op2_reg);
799 vfp_liveness live = vfp_live_args_after(sim, n, REGMASK(out));
803 DB((dbg, LEVEL_1, ">>> %+F %s, %s -> %s\n", n,
804 arch_register_get_name(op1_reg), arch_register_get_name(op2_reg),
805 arch_register_get_name(out)));
806 DEBUG_ONLY(vfp_dump_live(live);)
807 DB((dbg, LEVEL_1, "Stack before: "));
808 DEBUG_ONLY(x87_dump_stack(state);)
810 op1_idx = x87_on_stack(state, reg_index_1);
811 assert(op1_idx >= 0);
812 op1_live_after = is_vfp_live(reg_index_1, live);
814 attr = get_ia32_x87_attr(n);
815 permuted = attr->attr.data.ins_permuted;
817 if (reg_index_2 != REG_VFP_VFP_NOREG) {
820 /* second operand is a vfp register */
821 op2_idx = x87_on_stack(state, reg_index_2);
822 assert(op2_idx >= 0);
823 op2_live_after = is_vfp_live(reg_index_2, live);
825 if (op2_live_after) {
826 /* Second operand is live. */
828 if (op1_live_after) {
829 /* Both operands are live: push the first one.
830 This works even for op1 == op2. */
831 x87_create_fpush(state, n, op1_idx, op2);
832 /* now do fxxx (tos=tos X op) */
836 dst = tmpl->normal_op;
838 /* Second live, first operand is dead here, bring it to tos. */
840 x87_create_fxch(state, n, op1_idx);
845 /* now do fxxx (tos=tos X op) */
847 dst = tmpl->normal_op;
850 /* Second operand is dead. */
851 if (op1_live_after) {
852 /* First operand is live: bring second to tos. */
854 x87_create_fxch(state, n, op2_idx);
859 /* now do fxxxr (tos = op X tos) */
861 dst = tmpl->reverse_op;
863 /* Both operands are dead here, pop them from the stack. */
866 /* Both are identically and on tos, no pop needed. */
867 /* here fxxx (tos = tos X tos) */
868 dst = tmpl->normal_op;
871 /* now do fxxxp (op = op X tos, pop) */
872 dst = tmpl->normal_pop_op;
876 } else if (op1_idx == 0) {
877 assert(op1_idx != op2_idx);
878 /* now do fxxxrp (op = tos X op, pop) */
879 dst = tmpl->reverse_pop_op;
883 /* Bring the second on top. */
884 x87_create_fxch(state, n, op2_idx);
885 if (op1_idx == op2_idx) {
886 /* Both are identically and on tos now, no pop needed. */
889 /* use fxxx (tos = tos X tos) */
890 dst = tmpl->normal_op;
893 /* op2 is on tos now */
895 /* use fxxxp (op = op X tos, pop) */
896 dst = tmpl->normal_pop_op;
904 /* second operand is an address mode */
905 if (op1_live_after) {
906 /* first operand is live: push it here */
907 x87_create_fpush(state, n, op1_idx, op1);
910 /* first operand is dead: bring it to tos */
912 x87_create_fxch(state, n, op1_idx);
917 /* use fxxx (tos = tos X mem) */
918 dst = permuted ? tmpl->reverse_op : tmpl->normal_op;
922 patched_insn = x87_patch_insn(n, dst);
923 x87_set_st(state, arch_register_get_index(out), patched_insn, out_idx);
928 /* patch the operation */
929 attr->x87[0] = op1_reg = get_st_reg(op1_idx);
930 if (reg_index_2 != REG_VFP_VFP_NOREG) {
931 attr->x87[1] = op2_reg = get_st_reg(op2_idx);
933 attr->x87[2] = out = get_st_reg(out_idx);
935 if (reg_index_2 != REG_VFP_VFP_NOREG) {
936 DB((dbg, LEVEL_1, "<<< %s %s, %s -> %s\n", get_irn_opname(n),
937 arch_register_get_name(op1_reg), arch_register_get_name(op2_reg),
938 arch_register_get_name(out)));
940 DB((dbg, LEVEL_1, "<<< %s %s, [AM] -> %s\n", get_irn_opname(n),
941 arch_register_get_name(op1_reg),
942 arch_register_get_name(out)));
945 return NO_NODE_ADDED;
949 * Simulate a virtual Unop.
951 * @param state the x87 state
952 * @param n the node that should be simulated (and patched)
953 * @param op the x87 opcode that will replace n's opcode
955 * @return NO_NODE_ADDED
957 static int sim_unop(x87_state *state, ir_node *n, ir_op *op)
959 arch_register_t const *const out = x87_get_irn_register(n);
960 unsigned const live = vfp_live_args_after(state->sim, n, REGMASK(out));
961 DB((dbg, LEVEL_1, ">>> %+F -> %s\n", n, out->name));
962 DEBUG_ONLY(vfp_dump_live(live);)
964 ir_node *const op1 = get_irn_n(n, 0);
965 arch_register_t const *const op1_reg = x87_get_irn_register(op1);
966 int const op1_reg_idx = arch_register_get_index(op1_reg);
967 int const op1_idx = x87_on_stack(state, op1_reg_idx);
968 if (is_vfp_live(op1_reg_idx, live)) {
969 /* push the operand here */
970 x87_create_fpush(state, n, op1_idx, op1);
972 /* operand is dead, bring it to tos */
974 x87_create_fxch(state, n, op1_idx);
978 x87_set_tos(state, arch_register_get_index(out), x87_patch_insn(n, op));
979 ia32_x87_attr_t *const attr = get_ia32_x87_attr(n);
980 attr->x87[2] = attr->x87[0] = get_st_reg(0);
981 DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), attr->x87[2]->name));
983 return NO_NODE_ADDED;
987 * Simulate a virtual Load instruction.
989 * @param state the x87 state
990 * @param n the node that should be simulated (and patched)
991 * @param op the x87 opcode that will replace n's opcode
993 * @return NO_NODE_ADDED
995 static int sim_load(x87_state *state, ir_node *n, ir_op *op, int res_pos)
997 const arch_register_t *out = x87_irn_get_register(n, res_pos);
998 ia32_x87_attr_t *attr;
1000 DB((dbg, LEVEL_1, ">>> %+F -> %s\n", n, arch_register_get_name(out)));
1001 x87_push(state, arch_register_get_index(out), x87_patch_insn(n, op));
1002 assert(out == x87_irn_get_register(n, res_pos));
1003 attr = get_ia32_x87_attr(n);
1004 attr->x87[2] = out = get_st_reg(0);
1005 DB((dbg, LEVEL_1, "<<< %s -> %s\n", get_irn_opname(n), arch_register_get_name(out)));
1007 return NO_NODE_ADDED;
1011 * Rewire all users of @p old_val to @new_val iff they are scheduled after @p store.
1013 * @param store The store
1014 * @param old_val The former value
1015 * @param new_val The new value
1017 static void collect_and_rewire_users(ir_node *store, ir_node *old_val, ir_node *new_val)
1019 foreach_out_edge_safe(old_val, edge) {
1020 ir_node *user = get_edge_src_irn(edge);
1022 if (! user || user == store)
1025 /* if the user is scheduled after the store: rewire */
1026 if (sched_is_scheduled(user) && sched_comes_after(store, user)) {
1027 set_irn_n(user, get_edge_src_pos(edge), new_val);
1033 * Simulate a virtual Store.
1035 * @param state the x87 state
1036 * @param n the node that should be simulated (and patched)
1037 * @param op the x87 store opcode
1038 * @param op_p the x87 store and pop opcode
1040 static int sim_store(x87_state *state, ir_node *n, ir_op *op, ir_op *op_p)
1042 ir_node *const val = get_irn_n(n, n_ia32_vfst_val);
1043 arch_register_t const *const op2 = x87_get_irn_register(val);
1044 DB((dbg, LEVEL_1, ">>> %+F %s ->\n", n, arch_register_get_name(op2)));
1046 int insn = NO_NODE_ADDED;
1047 int const op2_reg_idx = arch_register_get_index(op2);
1048 int const op2_idx = x87_on_stack(state, op2_reg_idx);
1049 unsigned const live = vfp_live_args_after(state->sim, n, 0);
1050 int const live_after_node = is_vfp_live(op2_reg_idx, live);
1051 assert(op2_idx >= 0);
1052 if (live_after_node) {
1053 /* Problem: fst doesn't support 80bit modes (spills), only fstp does
1054 * fist doesn't support 64bit mode, only fistp
1056 * - stack not full: push value and fstp
1057 * - stack full: fstp value and load again
1058 * Note that we cannot test on mode_E, because floats might be 80bit ... */
1059 ir_mode *const mode = get_ia32_ls_mode(n);
1060 if (get_mode_size_bits(mode) > (mode_is_int(mode) ? 32 : 64)) {
1061 if (x87_get_depth(state) < N_ia32_st_REGS) {
1062 /* ok, we have a free register: push + fstp */
1063 x87_create_fpush(state, n, op2_idx, val);
1065 x87_patch_insn(n, op_p);
1067 /* stack full here: need fstp + load */
1069 x87_patch_insn(n, op_p);
1071 ir_node *const block = get_nodes_block(n);
1072 ir_graph *const irg = get_irn_irg(n);
1073 ir_node *const nomem = get_irg_no_mem(irg);
1074 ir_node *const vfld = new_bd_ia32_vfld(NULL, block, get_irn_n(n, 0), get_irn_n(n, 1), nomem, mode);
1076 /* copy all attributes */
1077 set_ia32_frame_ent(vfld, get_ia32_frame_ent(n));
1078 if (is_ia32_use_frame(n))
1079 set_ia32_use_frame(vfld);
1080 set_ia32_op_type(vfld, ia32_AddrModeS);
1081 add_ia32_am_offs_int(vfld, get_ia32_am_offs_int(n));
1082 set_ia32_am_sc(vfld, get_ia32_am_sc(n));
1083 set_ia32_ls_mode(vfld, mode);
1085 ir_node *const rproj = new_r_Proj(vfld, mode, pn_ia32_vfld_res);
1086 ir_node *const mproj = new_r_Proj(vfld, mode_M, pn_ia32_vfld_M);
1087 ir_node *const mem = get_irn_Proj_for_mode(n, mode_M);
1089 assert(mem && "Store memory not found");
1091 arch_set_irn_register(rproj, op2);
1093 /* reroute all former users of the store memory to the load memory */
1094 edges_reroute(mem, mproj);
1095 /* set the memory input of the load to the store memory */
1096 set_irn_n(vfld, n_ia32_vfld_mem, mem);
1098 sched_add_after(n, vfld);
1099 sched_add_after(vfld, rproj);
1101 /* rewire all users, scheduled after the store, to the loaded value */
1102 collect_and_rewire_users(n, val, rproj);
1107 /* we can only store the tos to memory */
1109 x87_create_fxch(state, n, op2_idx);
1111 /* mode size 64 or smaller -> use normal fst */
1112 x87_patch_insn(n, op);
1115 /* we can only store the tos to memory */
1117 x87_create_fxch(state, n, op2_idx);
1120 x87_patch_insn(n, op_p);
1123 ia32_x87_attr_t *const attr = get_ia32_x87_attr(n);
1124 attr->x87[1] = get_st_reg(0);
1125 DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(attr->x87[1])));
1130 #define _GEN_BINOP(op, rev) \
1131 static int sim_##op(x87_state *state, ir_node *n) { \
1132 exchange_tmpl tmpl = { op_ia32_##op, op_ia32_##rev, op_ia32_##op##p, op_ia32_##rev##p }; \
1133 return sim_binop(state, n, &tmpl); \
1136 #define GEN_BINOP(op) _GEN_BINOP(op, op)
1137 #define GEN_BINOPR(op) _GEN_BINOP(op, op##r)
1139 #define GEN_LOAD(op) \
1140 static int sim_##op(x87_state *state, ir_node *n) { \
1141 return sim_load(state, n, op_ia32_##op, pn_ia32_v##op##_res); \
1144 #define GEN_UNOP(op) \
1145 static int sim_##op(x87_state *state, ir_node *n) { \
1146 return sim_unop(state, n, op_ia32_##op); \
1149 #define GEN_STORE(op) \
1150 static int sim_##op(x87_state *state, ir_node *n) { \
1151 return sim_store(state, n, op_ia32_##op, op_ia32_##op##p); \
1173 * Simulate a virtual fisttp.
1175 * @param state the x87 state
1176 * @param n the node that should be simulated (and patched)
1178 * @return NO_NODE_ADDED
1180 static int sim_fisttp(x87_state *state, ir_node *n)
1182 ir_node *val = get_irn_n(n, n_ia32_vfst_val);
1183 const arch_register_t *op2 = x87_get_irn_register(val);
1184 ia32_x87_attr_t *attr;
1185 int op2_reg_idx, op2_idx;
1187 op2_reg_idx = arch_register_get_index(op2);
1188 op2_idx = x87_on_stack(state, op2_reg_idx);
1189 DB((dbg, LEVEL_1, ">>> %+F %s ->\n", n, arch_register_get_name(op2)));
1190 assert(op2_idx >= 0);
1192 /* Note: although the value is still live here, it is destroyed because
1193 of the pop. The register allocator is aware of that and introduced a copy
1194 if the value must be alive. */
1196 /* we can only store the tos to memory */
1198 x87_create_fxch(state, n, op2_idx);
1201 x87_patch_insn(n, op_ia32_fisttp);
1203 attr = get_ia32_x87_attr(n);
1204 attr->x87[1] = op2 = get_st_reg(0);
1205 DB((dbg, LEVEL_1, "<<< %s %s ->\n", get_irn_opname(n), arch_register_get_name(op2)));
1207 return NO_NODE_ADDED;
1211 * Simulate a virtual FtstFnstsw.
1213 * @param state the x87 state
1214 * @param n the node that should be simulated (and patched)
1216 * @return NO_NODE_ADDED
1218 static int sim_FtstFnstsw(x87_state *state, ir_node *n)
1220 x87_simulator *sim = state->sim;
1221 ia32_x87_attr_t *attr = get_ia32_x87_attr(n);
1222 ir_node *op1_node = get_irn_n(n, n_ia32_vFtstFnstsw_left);
1223 const arch_register_t *reg1 = x87_get_irn_register(op1_node);
1224 int reg_index_1 = arch_register_get_index(reg1);
1225 int op1_idx = x87_on_stack(state, reg_index_1);
1226 unsigned live = vfp_live_args_after(sim, n, 0);
1228 DB((dbg, LEVEL_1, ">>> %+F %s\n", n, arch_register_get_name(reg1)));
1229 DEBUG_ONLY(vfp_dump_live(live);)
1230 DB((dbg, LEVEL_1, "Stack before: "));
1231 DEBUG_ONLY(x87_dump_stack(state);)
1232 assert(op1_idx >= 0);
1235 /* bring the value to tos */
1236 x87_create_fxch(state, n, op1_idx);
1240 /* patch the operation */
1241 x87_patch_insn(n, op_ia32_FtstFnstsw);
1242 reg1 = get_st_reg(op1_idx);
1243 attr->x87[0] = reg1;
1244 attr->x87[1] = NULL;
1245 attr->x87[2] = NULL;
1247 if (!is_vfp_live(reg_index_1, live))
1248 x87_create_fpop(state, sched_next(n), 1);
1250 return NO_NODE_ADDED;
1256 * @param state the x87 state
1257 * @param n the node that should be simulated (and patched)
1259 * @return NO_NODE_ADDED
1261 static int sim_Fucom(x87_state *state, ir_node *n)
1265 ia32_x87_attr_t *attr = get_ia32_x87_attr(n);
1267 x87_simulator *sim = state->sim;
1268 ir_node *op1_node = get_irn_n(n, n_ia32_vFucomFnstsw_left);
1269 ir_node *op2_node = get_irn_n(n, n_ia32_vFucomFnstsw_right);
1270 const arch_register_t *op1 = x87_get_irn_register(op1_node);
1271 const arch_register_t *op2 = x87_get_irn_register(op2_node);
1272 int reg_index_1 = arch_register_get_index(op1);
1273 int reg_index_2 = arch_register_get_index(op2);
1274 unsigned live = vfp_live_args_after(sim, n, 0);
1275 bool permuted = attr->attr.data.ins_permuted;
1279 DB((dbg, LEVEL_1, ">>> %+F %s, %s\n", n,
1280 arch_register_get_name(op1), arch_register_get_name(op2)));
1281 DEBUG_ONLY(vfp_dump_live(live);)
1282 DB((dbg, LEVEL_1, "Stack before: "));
1283 DEBUG_ONLY(x87_dump_stack(state);)
1285 op1_idx = x87_on_stack(state, reg_index_1);
1286 assert(op1_idx >= 0);
1288 /* BEWARE: check for comp a,a cases, they might happen */
1289 if (reg_index_2 != REG_VFP_VFP_NOREG) {
1290 /* second operand is a vfp register */
1291 op2_idx = x87_on_stack(state, reg_index_2);
1292 assert(op2_idx >= 0);
1294 if (is_vfp_live(reg_index_2, live)) {
1295 /* second operand is live */
1297 if (is_vfp_live(reg_index_1, live)) {
1298 /* both operands are live */
1301 /* res = tos X op */
1302 } else if (op2_idx == 0) {
1303 /* res = op X tos */
1304 permuted = !permuted;
1307 /* bring the first one to tos */
1308 x87_create_fxch(state, n, op1_idx);
1309 if (op1_idx == op2_idx) {
1311 } else if (op2_idx == 0) {
1315 /* res = tos X op */
1318 /* second live, first operand is dead here, bring it to tos.
1319 This means further, op1_idx != op2_idx. */
1320 assert(op1_idx != op2_idx);
1322 x87_create_fxch(state, n, op1_idx);
1327 /* res = tos X op, pop */
1331 /* second operand is dead */
1332 if (is_vfp_live(reg_index_1, live)) {
1333 /* first operand is live: bring second to tos.
1334 This means further, op1_idx != op2_idx. */
1335 assert(op1_idx != op2_idx);
1337 x87_create_fxch(state, n, op2_idx);
1342 /* res = op X tos, pop */
1344 permuted = !permuted;
1347 /* both operands are dead here, check first for identity. */
1348 if (op1_idx == op2_idx) {
1349 /* identically, one pop needed */
1351 x87_create_fxch(state, n, op1_idx);
1355 /* res = tos X op, pop */
1358 /* different, move them to st and st(1) and pop both.
1359 The tricky part is to get one into st(1).*/
1360 else if (op2_idx == 1) {
1361 /* good, second operand is already in the right place, move the first */
1363 /* bring the first on top */
1364 x87_create_fxch(state, n, op1_idx);
1365 assert(op2_idx != 0);
1368 /* res = tos X op, pop, pop */
1370 } else if (op1_idx == 1) {
1371 /* good, first operand is already in the right place, move the second */
1373 /* bring the first on top */
1374 x87_create_fxch(state, n, op2_idx);
1375 assert(op1_idx != 0);
1378 /* res = op X tos, pop, pop */
1379 permuted = !permuted;
1383 /* if one is already the TOS, we need two fxch */
1385 /* first one is TOS, move to st(1) */
1386 x87_create_fxch(state, n, 1);
1387 assert(op2_idx != 1);
1389 x87_create_fxch(state, n, op2_idx);
1391 /* res = op X tos, pop, pop */
1393 permuted = !permuted;
1395 } else if (op2_idx == 0) {
1396 /* second one is TOS, move to st(1) */
1397 x87_create_fxch(state, n, 1);
1398 assert(op1_idx != 1);
1400 x87_create_fxch(state, n, op1_idx);
1402 /* res = tos X op, pop, pop */
1405 /* none of them is either TOS or st(1), 3 fxch needed */
1406 x87_create_fxch(state, n, op2_idx);
1407 assert(op1_idx != 0);
1408 x87_create_fxch(state, n, 1);
1410 x87_create_fxch(state, n, op1_idx);
1412 /* res = tos X op, pop, pop */
1419 /* second operand is an address mode */
1420 if (is_vfp_live(reg_index_1, live)) {
1421 /* first operand is live: bring it to TOS */
1423 x87_create_fxch(state, n, op1_idx);
1427 /* first operand is dead: bring it to tos */
1429 x87_create_fxch(state, n, op1_idx);
1436 /* patch the operation */
1437 if (is_ia32_vFucomFnstsw(n)) {
1441 case 0: dst = op_ia32_FucomFnstsw; break;
1442 case 1: dst = op_ia32_FucompFnstsw; break;
1443 case 2: dst = op_ia32_FucomppFnstsw; break;
1444 default: panic("invalid popcount");
1447 for (i = 0; i < pops; ++i) {
1450 } else if (is_ia32_vFucomi(n)) {
1452 case 0: dst = op_ia32_Fucomi; break;
1453 case 1: dst = op_ia32_Fucompi; x87_pop(state); break;
1455 dst = op_ia32_Fucompi;
1457 x87_create_fpop(state, sched_next(n), 1);
1459 default: panic("invalid popcount");
1462 panic("invalid operation %+F", n);
1465 x87_patch_insn(n, dst);
1472 op1 = get_st_reg(op1_idx);
1475 op2 = get_st_reg(op2_idx);
1478 attr->x87[2] = NULL;
1479 attr->attr.data.ins_permuted = permuted;
1482 DB((dbg, LEVEL_1, "<<< %s %s, %s\n", get_irn_opname(n),
1483 arch_register_get_name(op1), arch_register_get_name(op2)));
1485 DB((dbg, LEVEL_1, "<<< %s %s, [AM]\n", get_irn_opname(n),
1486 arch_register_get_name(op1)));
1489 return NO_NODE_ADDED;
1495 * @param state the x87 state
1496 * @param n the node that should be simulated (and patched)
1498 * @return NO_NODE_ADDED
1500 static int sim_Keep(x87_state *state, ir_node *node)
1503 const arch_register_t *op_reg;
1509 DB((dbg, LEVEL_1, ">>> %+F\n", node));
1511 arity = get_irn_arity(node);
1512 for (i = 0; i < arity; ++i) {
1513 op = get_irn_n(node, i);
1514 op_reg = arch_get_irn_register(op);
1515 if (arch_register_get_class(op_reg) != &ia32_reg_classes[CLASS_ia32_vfp])
1518 reg_id = arch_register_get_index(op_reg);
1519 live = vfp_live_args_after(state->sim, node, 0);
1521 op_stack_idx = x87_on_stack(state, reg_id);
1522 if (op_stack_idx >= 0 && !is_vfp_live(reg_id, live))
1523 x87_create_fpop(state, sched_next(node), 1);
1526 DB((dbg, LEVEL_1, "Stack after: "));
1527 DEBUG_ONLY(x87_dump_stack(state);)
1529 return NO_NODE_ADDED;
1533 * Keep the given node alive by adding a be_Keep.
1535 * @param node the node to kept alive
1537 static void keep_float_node_alive(ir_node *node)
1539 ir_node *block = get_nodes_block(node);
1540 ir_node *keep = be_new_Keep(block, 1, &node);
1542 assert(sched_is_scheduled(node));
1543 sched_add_after(node, keep);
1547 * Create a copy of a node. Recreate the node if it's a constant.
1549 * @param state the x87 state
1550 * @param n the node to be copied
1552 * @return the copy of n
1554 static ir_node *create_Copy(x87_state *state, ir_node *n)
1556 dbg_info *n_dbg = get_irn_dbg_info(n);
1557 ir_mode *mode = get_irn_mode(n);
1558 ir_node *block = get_nodes_block(n);
1559 ir_node *pred = get_irn_n(n, 0);
1560 ir_node *(*cnstr)(dbg_info *, ir_node *, ir_mode *) = NULL;
1562 const arch_register_t *out;
1563 const arch_register_t *op1;
1564 ia32_x87_attr_t *attr;
1566 /* Do not copy constants, recreate them. */
1567 switch (get_ia32_irn_opcode(pred)) {
1569 cnstr = new_bd_ia32_fldz;
1572 cnstr = new_bd_ia32_fld1;
1574 case iro_ia32_fldpi:
1575 cnstr = new_bd_ia32_fldpi;
1577 case iro_ia32_fldl2e:
1578 cnstr = new_bd_ia32_fldl2e;
1580 case iro_ia32_fldl2t:
1581 cnstr = new_bd_ia32_fldl2t;
1583 case iro_ia32_fldlg2:
1584 cnstr = new_bd_ia32_fldlg2;
1586 case iro_ia32_fldln2:
1587 cnstr = new_bd_ia32_fldln2;
1593 out = x87_get_irn_register(n);
1594 op1 = x87_get_irn_register(pred);
1596 if (cnstr != NULL) {
1597 /* copy a constant */
1598 res = (*cnstr)(n_dbg, block, mode);
1600 x87_push(state, arch_register_get_index(out), res);
1602 attr = get_ia32_x87_attr(res);
1603 attr->x87[2] = get_st_reg(0);
1605 int op1_idx = x87_on_stack(state, arch_register_get_index(op1));
1607 res = new_bd_ia32_fpushCopy(n_dbg, block, pred, mode);
1609 x87_push(state, arch_register_get_index(out), res);
1611 attr = get_ia32_x87_attr(res);
1612 attr->x87[0] = get_st_reg(op1_idx);
1613 attr->x87[2] = get_st_reg(0);
1615 arch_set_irn_register(res, out);
1621 * Simulate a be_Copy.
1623 * @param state the x87 state
1624 * @param n the node that should be simulated (and patched)
1626 * @return NO_NODE_ADDED
1628 static int sim_Copy(x87_state *state, ir_node *n)
1630 arch_register_class_t const *const cls = arch_get_irn_reg_class(n);
1631 if (cls != &ia32_reg_classes[CLASS_ia32_vfp])
1632 return NO_NODE_ADDED;
1634 ir_node *const pred = be_get_Copy_op(n);
1635 arch_register_t const *const op1 = x87_get_irn_register(pred);
1636 arch_register_t const *const out = x87_get_irn_register(n);
1637 unsigned const live = vfp_live_args_after(state->sim, n, REGMASK(out));
1639 DB((dbg, LEVEL_1, ">>> %+F %s -> %s\n", n,
1640 arch_register_get_name(op1), arch_register_get_name(out)));
1641 DEBUG_ONLY(vfp_dump_live(live);)
1643 if (is_vfp_live(arch_register_get_index(op1), live)) {
1644 /* Operand is still live, a real copy. We need here an fpush that can
1645 hold a a register, so use the fpushCopy or recreate constants */
1646 ir_node *const node = create_Copy(state, n);
1648 /* We have to make sure the old value doesn't go dead (which can happen
1649 * when we recreate constants). As the simulator expected that value in
1650 * the pred blocks. This is unfortunate as removing it would save us 1
1651 * instruction, but we would have to rerun all the simulation to get
1654 ir_node *const next = sched_next(n);
1657 sched_add_before(next, node);
1659 if (get_irn_n_edges(pred) == 0) {
1660 keep_float_node_alive(pred);
1663 DB((dbg, LEVEL_1, "<<< %+F %s -> ?\n", node, op1->name));
1665 int const op1_idx = x87_on_stack(state, arch_register_get_index(op1));
1666 int const out_idx = x87_on_stack(state, arch_register_get_index(out));
1667 if (out_idx >= 0 && out_idx != op1_idx) {
1668 /* Matze: out already on stack? how can this happen? */
1669 panic("invalid stack state");
1672 /* op1 must be killed and placed where out is */
1674 ia32_x87_attr_t *attr;
1675 /* best case, simple remove and rename */
1676 x87_patch_insn(n, op_ia32_Pop);
1677 attr = get_ia32_x87_attr(n);
1678 attr->x87[0] = op1 = get_st_reg(0);
1681 x87_set_st(state, arch_register_get_index(out), n, op1_idx - 1);
1683 ia32_x87_attr_t *attr;
1684 /* move op1 to tos, store and pop it */
1686 x87_create_fxch(state, n, op1_idx);
1689 x87_patch_insn(n, op_ia32_Pop);
1690 attr = get_ia32_x87_attr(n);
1691 attr->x87[0] = op1 = get_st_reg(out_idx);
1694 x87_set_st(state, arch_register_get_index(out), n, out_idx - 1);
1696 DB((dbg, LEVEL_1, "<<< %+F %s\n", n, op1->name));
1699 /* just a virtual copy */
1700 x87_set_st(state, arch_register_get_index(out), pred, op1_idx);
1701 /* don't remove the node to keep the verifier quiet :),
1702 the emitter won't emit any code for the node */
1705 DB((dbg, LEVEL_1, "<<< KILLED %s\n", get_irn_opname(n)));
1710 return NO_NODE_ADDED;
1714 * Returns the vf0 result Proj of a Call.
1716 * @para call the Call node
1718 static ir_node *get_call_result_proj(ir_node *call)
1720 /* search the result proj */
1721 foreach_out_edge(call, edge) {
1722 ir_node *proj = get_edge_src_irn(edge);
1723 long pn = get_Proj_proj(proj);
1725 if (pn == pn_ia32_Call_vf0)
1729 panic("result Proj missing");
1732 static int sim_Asm(x87_state *const state, ir_node *const n)
1736 for (size_t i = get_irn_arity(n); i-- != 0;) {
1737 arch_register_req_t const *const req = arch_get_irn_register_req_in(n, i);
1738 if (req->cls == &ia32_reg_classes[CLASS_ia32_vfp])
1739 panic("cannot handle %+F with x87 constraints", n);
1742 for (size_t i = arch_get_irn_n_outs(n); i-- != 0;) {
1743 arch_register_req_t const *const req = arch_get_irn_register_req_out(n, i);
1744 if (req->cls == &ia32_reg_classes[CLASS_ia32_vfp])
1745 panic("cannot handle %+F with x87 constraints", n);
1748 return NO_NODE_ADDED;
1752 * Simulate a ia32_Call.
1754 * @param state the x87 state
1755 * @param n the node that should be simulated (and patched)
1757 * @return NO_NODE_ADDED
1759 static int sim_Call(x87_state *state, ir_node *n)
1761 DB((dbg, LEVEL_1, ">>> %+F\n", n));
1763 /* at the begin of a call the x87 state should be empty */
1764 assert(state->depth == 0 && "stack not empty before call");
1766 ir_type *const call_tp = get_ia32_call_attr_const(n)->call_tp;
1767 if (get_method_n_ress(call_tp) != 0) {
1768 /* If the called function returns a float, it is returned in st(0).
1769 * This even happens if the return value is NOT used.
1770 * Moreover, only one return result is supported. */
1771 ir_type *const res_type = get_method_res_type(call_tp, 0);
1772 ir_mode *const mode = get_type_mode(res_type);
1773 if (mode && mode_is_float(mode)) {
1774 ir_node *const resproj = get_call_result_proj(n);
1775 arch_register_t const *const reg = x87_get_irn_register(resproj);
1776 x87_push(state, arch_register_get_index(reg), resproj);
1779 DB((dbg, LEVEL_1, "Stack after: "));
1780 DEBUG_ONLY(x87_dump_stack(state);)
1782 return NO_NODE_ADDED;
1786 * Simulate a be_Return.
1788 * @param state the x87 state
1789 * @param n the node that should be simulated (and patched)
1791 * @return NO_NODE_ADDED
1793 static int sim_Return(x87_state *state, ir_node *n)
1795 #ifdef DEBUG_libfirm
1796 /* only floating point return values must reside on stack */
1797 int n_float_res = 0;
1798 int const n_res = be_Return_get_n_rets(n);
1799 for (int i = 0; i < n_res; ++i) {
1800 ir_node *const res = get_irn_n(n, n_be_Return_val + i);
1801 if (mode_is_float(get_irn_mode(res)))
1804 assert(x87_get_depth(state) == n_float_res);
1807 /* pop them virtually */
1809 return NO_NODE_ADDED;
1813 * Simulate a be_Perm.
1815 * @param state the x87 state
1816 * @param irn the node that should be simulated (and patched)
1818 * @return NO_NODE_ADDED
1820 static int sim_Perm(x87_state *state, ir_node *irn)
1823 ir_node *pred = get_irn_n(irn, 0);
1826 /* handle only floating point Perms */
1827 if (! mode_is_float(get_irn_mode(pred)))
1828 return NO_NODE_ADDED;
1830 DB((dbg, LEVEL_1, ">>> %+F\n", irn));
1832 /* Perm is a pure virtual instruction on x87.
1833 All inputs must be on the FPU stack and are pairwise
1834 different from each other.
1835 So, all we need to do is to permutate the stack state. */
1836 n = get_irn_arity(irn);
1837 NEW_ARR_A(int, stack_pos, n);
1839 /* collect old stack positions */
1840 for (i = 0; i < n; ++i) {
1841 const arch_register_t *inreg = x87_get_irn_register(get_irn_n(irn, i));
1842 int idx = x87_on_stack(state, arch_register_get_index(inreg));
1844 assert(idx >= 0 && "Perm argument not on x87 stack");
1848 /* now do the permutation */
1849 foreach_out_edge(irn, edge) {
1850 ir_node *proj = get_edge_src_irn(edge);
1851 const arch_register_t *out = x87_get_irn_register(proj);
1852 long num = get_Proj_proj(proj);
1854 assert(0 <= num && num < n && "More Proj's than Perm inputs");
1855 x87_set_st(state, arch_register_get_index(out), proj, stack_pos[(unsigned)num]);
1857 DB((dbg, LEVEL_1, "<<< %+F\n", irn));
1859 return NO_NODE_ADDED;
1863 * Kill any dead registers at block start by popping them from the stack.
1865 * @param sim the simulator handle
1866 * @param block the current block
1867 * @param state the x87 state at the begin of the block
1869 static void x87_kill_deads(x87_simulator *const sim, ir_node *const block, x87_state *const state)
1871 ir_node *first_insn = sched_first(block);
1872 ir_node *keep = NULL;
1873 unsigned live = vfp_live_args_after(sim, block, 0);
1875 int i, depth, num_pop;
1878 depth = x87_get_depth(state);
1879 for (i = depth - 1; i >= 0; --i) {
1880 int reg = x87_get_st_reg(state, i);
1882 if (! is_vfp_live(reg, live))
1883 kill_mask |= (1 << i);
1887 DB((dbg, LEVEL_1, "Killing deads:\n"));
1888 DEBUG_ONLY(vfp_dump_live(live);)
1889 DEBUG_ONLY(x87_dump_stack(state);)
1891 if (kill_mask != 0 && live == 0) {
1892 /* special case: kill all registers */
1893 if (ia32_cg_config.use_femms || ia32_cg_config.use_emms) {
1894 if (ia32_cg_config.use_femms) {
1895 /* use FEMMS on AMD processors to clear all */
1896 keep = new_bd_ia32_femms(NULL, block);
1898 /* use EMMS to clear all */
1899 keep = new_bd_ia32_emms(NULL, block);
1901 sched_add_before(first_insn, keep);
1907 /* now kill registers */
1909 /* we can only kill from TOS, so bring them up */
1910 if (! (kill_mask & 1)) {
1911 /* search from behind, because we can to a double-pop */
1912 for (i = depth - 1; i >= 0; --i) {
1913 if (kill_mask & (1 << i)) {
1914 kill_mask &= ~(1 << i);
1921 x87_set_st(state, -1, keep, i);
1922 x87_create_fxch(state, first_insn, i);
1925 if ((kill_mask & 3) == 3) {
1926 /* we can do a double-pop */
1930 /* only a single pop */
1935 kill_mask >>= num_pop;
1936 keep = x87_create_fpop(state, first_insn, num_pop);
1943 * Run a simulation and fix all virtual instructions for a block.
1945 * @param sim the simulator handle
1946 * @param block the current block
1948 static void x87_simulate_block(x87_simulator *sim, ir_node *block)
1951 blk_state *bl_state = x87_get_bl_state(sim, block);
1952 x87_state *state = bl_state->begin;
1953 ir_node *start_block;
1955 assert(state != NULL);
1956 /* already processed? */
1957 if (bl_state->end != NULL)
1960 DB((dbg, LEVEL_1, "Simulate %+F\n", block));
1961 DB((dbg, LEVEL_2, "State at Block begin:\n "));
1962 DEBUG_ONLY(x87_dump_stack(state);)
1964 /* create a new state, will be changed */
1965 state = x87_clone_state(sim, state);
1966 /* at block begin, kill all dead registers */
1967 x87_kill_deads(sim, block, state);
1969 /* beware, n might change */
1970 for (n = sched_first(block); !sched_is_end(n); n = next) {
1973 ir_op *op = get_irn_op(n);
1976 * get the next node to be simulated here.
1977 * n might be completely removed from the schedule-
1979 next = sched_next(n);
1980 if (op->ops.generic != NULL) {
1981 func = (sim_func)op->ops.generic;
1984 node_inserted = (*func)(state, n);
1987 * sim_func might have added an additional node after n,
1988 * so update next node
1989 * beware: n must not be changed by sim_func
1990 * (i.e. removed from schedule) in this case
1992 if (node_inserted != NO_NODE_ADDED)
1993 next = sched_next(n);
1997 start_block = get_irg_start_block(get_irn_irg(block));
1999 DB((dbg, LEVEL_2, "State at Block end:\n ")); DEBUG_ONLY(x87_dump_stack(state);)
2001 /* check if the state must be shuffled */
2002 foreach_block_succ(block, edge) {
2003 ir_node *succ = get_edge_src_irn(edge);
2004 blk_state *succ_state;
2006 if (succ == start_block)
2009 succ_state = x87_get_bl_state(sim, succ);
2011 if (succ_state->begin == NULL) {
2012 DB((dbg, LEVEL_2, "Set begin state for succ %+F:\n", succ));
2013 DEBUG_ONLY(x87_dump_stack(state);)
2014 succ_state->begin = state;
2016 waitq_put(sim->worklist, succ);
2018 DB((dbg, LEVEL_2, "succ %+F already has a state, shuffling\n", succ));
2019 /* There is already a begin state for the successor, bad.
2020 Do the necessary permutations.
2021 Note that critical edges are removed, so this is always possible:
2022 If the successor has more than one possible input, then it must
2025 x87_shuffle(block, state, succ_state->begin);
2028 bl_state->end = state;
2032 * Register a simulator function.
2034 * @param op the opcode to simulate
2035 * @param func the simulator function for the opcode
2037 static void register_sim(ir_op *op, sim_func func)
2039 assert(op->ops.generic == NULL);
2040 op->ops.generic = (op_func) func;
2044 * Create a new x87 simulator.
2046 * @param sim a simulator handle, will be initialized
2047 * @param irg the current graph
2049 static void x87_init_simulator(x87_simulator *sim, ir_graph *irg)
2051 obstack_init(&sim->obst);
2052 sim->blk_states = pmap_create();
2053 sim->n_idx = get_irg_last_idx(irg);
2054 sim->live = OALLOCN(&sim->obst, vfp_liveness, sim->n_idx);
2056 DB((dbg, LEVEL_1, "--------------------------------\n"
2057 "x87 Simulator started for %+F\n", irg));
2059 /* set the generic function pointer of instruction we must simulate */
2060 ir_clear_opcodes_generic_func();
2062 register_sim(op_ia32_Asm, sim_Asm);
2063 register_sim(op_ia32_Call, sim_Call);
2064 register_sim(op_ia32_vfld, sim_fld);
2065 register_sim(op_ia32_vfild, sim_fild);
2066 register_sim(op_ia32_vfld1, sim_fld1);
2067 register_sim(op_ia32_vfldz, sim_fldz);
2068 register_sim(op_ia32_vfadd, sim_fadd);
2069 register_sim(op_ia32_vfsub, sim_fsub);
2070 register_sim(op_ia32_vfmul, sim_fmul);
2071 register_sim(op_ia32_vfdiv, sim_fdiv);
2072 register_sim(op_ia32_vfprem, sim_fprem);
2073 register_sim(op_ia32_vfabs, sim_fabs);
2074 register_sim(op_ia32_vfchs, sim_fchs);
2075 register_sim(op_ia32_vfist, sim_fist);
2076 register_sim(op_ia32_vfisttp, sim_fisttp);
2077 register_sim(op_ia32_vfst, sim_fst);
2078 register_sim(op_ia32_vFtstFnstsw, sim_FtstFnstsw);
2079 register_sim(op_ia32_vFucomFnstsw, sim_Fucom);
2080 register_sim(op_ia32_vFucomi, sim_Fucom);
2081 register_sim(op_be_Copy, sim_Copy);
2082 register_sim(op_be_Return, sim_Return);
2083 register_sim(op_be_Perm, sim_Perm);
2084 register_sim(op_be_Keep, sim_Keep);
2088 * Destroy a x87 simulator.
2090 * @param sim the simulator handle
2092 static void x87_destroy_simulator(x87_simulator *sim)
2094 pmap_destroy(sim->blk_states);
2095 obstack_free(&sim->obst, NULL);
2096 DB((dbg, LEVEL_1, "x87 Simulator stopped\n\n"));
2100 * Pre-block walker: calculate the liveness information for the block
2101 * and store it into the sim->live cache.
2103 static void update_liveness_walker(ir_node *block, void *data)
2105 x87_simulator *sim = (x87_simulator*)data;
2106 update_liveness(sim, block);
2110 * Run a simulation and fix all virtual instructions for a graph.
2111 * Replaces all virtual floating point instructions and registers
2114 void ia32_x87_simulate_graph(ir_graph *irg)
2116 /* TODO improve code quality (less executed fxch) by using execfreqs */
2118 ir_node *block, *start_block;
2119 blk_state *bl_state;
2122 /* create the simulator */
2123 x87_init_simulator(&sim, irg);
2125 start_block = get_irg_start_block(irg);
2126 bl_state = x87_get_bl_state(&sim, start_block);
2128 /* start with the empty state */
2130 bl_state->begin = ∅
2132 sim.worklist = new_waitq();
2133 waitq_put(sim.worklist, start_block);
2135 be_assure_live_sets(irg);
2136 sim.lv = be_get_irg_liveness(irg);
2138 /* Calculate the liveness for all nodes. We must precalculate this info,
2139 * because the simulator adds new nodes (possible before Phi nodes) which
2140 * would let a lazy calculation fail.
2141 * On the other hand we reduce the computation amount due to
2142 * precaching from O(n^2) to O(n) at the expense of O(n) cache memory.
2144 irg_block_walk_graph(irg, update_liveness_walker, NULL, &sim);
2148 block = (ir_node*)waitq_get(sim.worklist);
2149 x87_simulate_block(&sim, block);
2150 } while (! waitq_empty(sim.worklist));
2153 del_waitq(sim.worklist);
2154 x87_destroy_simulator(&sim);
2157 /* Initializes the x87 simulator. */
2158 void ia32_init_x87(void)
2160 FIRM_DBG_REGISTER(dbg, "firm.be.ia32.x87");